Agent范式演进：ToolFormer工具学习

import re # 模拟 ToolFormer: 用带 <TOOL>/<RESULT> 标记的语料训练一个小模型# 模型学会在特定词后触发工具调用, 推理时执行工具并把结果写回序列 CORPUS = [    "今天天气<TOOL>search(weather)</TOOL><RESULT>晴,25度</RESULT>适合出门散步",    "5加7等于<TOOL>calc(5+7)</TOOL><RESULT>12</RESULT>回答完毕",    "矩形面积<TOOL>calc(3*4)</TOOL><RESULT>12</RESULT>平方米",    "北京到上海<TOOL>search(flight)</TOOL><RESULT>高铁4.5小时</RESULT>非常方便",] SEARCH_DB = {    "search(weather)": "晴,25度",    "search(flight)": "高铁4.5小时",} TOKEN_RE = re.compile(r"<TOOL>[^<]*</TOOL>|<RESULT>[^<]*</RESULT>|[^<>]+")  def tokenize(sentence):    return [t for t in TOKEN_RE.findall(sentence) if t.strip()]  def strip_tag(token):    return token.replace("<TOOL>", "").replace("</TOOL>", "")  class ToolFormer:    def __init__(self, corpus):        self.corpus = corpus        self.bigram_table = {}    # 前词 -> 后标记 -> 次数        self.tool_triggers = {}   # 前词 -> 触发工具调用次数     def train(self):        for sent in self.corpus:            tokens = tokenize(sent)            for prev, nxt in zip(tokens[:-1], tokens[1:]):                key = strip_tag(prev)                if key not in self.bigram_table:                    self.bigram_table[key] = {}                self.bigram_table[key][nxt] = self.bigram_table[key].get(nxt, 0) + 1                if "<TOOL>" in nxt:                    self.tool_triggers[key] = self.tool_triggers.get(key, 0) + 1     def predict_next(self, word):        table = self.bigram_table.get(word, {})        if not table:            return None        return max(table, key=table.get)     def generate(self, start, max_len=6):        out = [start]        word = start        for _ in range(max_len):            nxt = self.predict_next(word)            if nxt is None:                break            out.append(nxt)            if "<RESULT>" in nxt:                break            word = strip_tag(nxt)        return out  def execute_tool(call_token):    """模拟工具执行: calc 真算, search 查表"""    if call_token.startswith("<TOOL>calc"):        expr = re.sub(r"^<TOOL>calc\((.*?)\)</TOOL>$", r"\1", call_token)        try:            return str(eval(expr))        except Exception:            return "ERROR"    if call_token.startswith("<TOOL>search"):        name = re.sub(r"^<TOOL>(search\(.*?\))</TOOL>$", r"\1", call_token)        return SEARCH_DB.get(name, "未找到")    return "未知工具"  def main():    model = ToolFormer(CORPUS)    model.train()    print("=== 工具触发词统计(学到何时用工具) ===")    for word, count in sorted(model.tool_triggers.items(), key=lambda x: -x[1]):        print("  {0} 之后触发工具调用 {1} 次".format(word, count))    print()    print("=== 推理: 生成带工具调用的序列 ===")    for start in ["5加7等于", "今天天气", "矩形面积"]:        seq = model.generate(start)        text = "".join(seq)        print("生成: {0}".format(text))        for token in seq:            if "<TOOL>" in token:                result = execute_tool(token)                text = text.replace(token, "[执行: {0}]".format(result))            if "<RESULT>" in token:                text = text.replace(token, "")        print("执行后: {0}".format(text))        print()  if __name__ == "__main__":    main()